Opportunistic Spectrum Access in Multiple Primary User Environments Under the Packet Collision Constraint

Size: px
Start display at page:

Download "Opportunistic Spectrum Access in Multiple Primary User Environments Under the Packet Collision Constraint"

Transcription

1 IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. X, NO. Y, 1 Oortunistic Sectrum Access in Multile Primary User Environments Under the Packet Collision Constraint Eric Jung, Student Member, IEEE, and Xin Liu, Member, IEEE Abstract Cognitive Radio (CR) technology has great otential to alleviate sectrum scarcity in wireless communications. It allows secondary users (SUs) to oortunistically access sectrum licensed by rimary users (PUs) while rotecting PU activity. The rotection of the PUs is central to the adotion of this technology, since no PU would accommodate SU access to its own detriment. In this aer, we consider an SU that must rotect multile PUs simultaneously. We focus on the PU acket collision robability as the rotection metric. The PUs are unslotted and may have different idle/busy time distributions and rotection requirements. Under general idle time distributions, we determine the form of the SU otimal access olicy, and identify two secial cases for which the comutation of the otimal olicy is significantly reduced. We also resent a simle algorithm to determine these olicies using rinciles of convex otimization theory. We then derive the otimal olicy for the same system when a SU has extra side information on PU activity. We evaluate the erformance of these olicies through simulation. Index Terms Cognitive Radio, Dynamic Sectrum Access, Sectrum Sharing. I. INTRODUCTION The static nature of sectrum regulatory olicy in the U.S. has led to an artificial scarcity of available sectrum. The FCC has estimated that 6% of the sectrum below 6 GHz is underutilized under the current allocation olicy [1]. Cognitive radio technology has been considered to mitigate this roblem [2]. It enables a secondary user to sense channel conditions and change its oerating characteristics to oortunistically access unoccuied rimary sectral bands. This new aradigm is tyically referred to as dynamic sectrum access (DSA). In the hierarchical model of dynamic sectrum access summarized in [3], users in the system are divided into a multitiered hierarchy where certain users have riority of channel access over others. Cognitive radio is conceived as a way for unlicensed secondary users (SU) to oortunistically access licensed sectral bands if rimary user (PU) activity is rotected from interference [4], [5]. This model is necessary because users will not agree to accommodate secondary networks to their own detriment. Therefore, a design imerative for a SU oortunistic access strategy is to minimize the SUs effect on PU transmissions. For examle, in the DARPA XG Manuscrit received July 23, 211. E. Jung is with the Deartment of Electrical and Comuter Engineering, University of California, Davis, 1 Shields Avenue, Davis, CA eajung@ucdavis.edu. X. Liu is with the Deartment of Comuter Science, University of California, Davis, 1 Shields Avenue, Davis, CA liu@cs.ucdavis.edu. Fig. 1. Tx(PU1) Rx(PU2) Rx(SU) Tx(SU) Rx(PU1) Tx(PU1) (a) SU interfering with two PUs that are satially nonoverlaing. Rx(PU2) Tx(PU1) Tx(SU) Channel 2 Channel 1 Rx(PU1) Channel 1,2 Tx(PU2) Rx(SU) (b) SU accessing 2 PU channels simultaneously. SU in environments with multile non-interfering PUs. roject [6], one of the three major test criteria in a cognitive radio rototye field test is to cause no harm [7]. Predictably, this is also one of the main bottlenecks of SU erformance. In this work, we consider acket collision robability as the PU rotection requirement. Under this requirement, the SU must guarantee that the acket collision robability of a PU acket is less than a certain threshold secified by the PU. This tye of constraint has already received some attention [8], [9], [1], [11]. In articular, we investigate SU erformance in a system with multile PUs with different acket collision robability constraints and usage atterns. There are common situations where multile PU systems may need to be rotected that have different owners, riorities, and usage atterns. Two simle examles are cases where the SU interferes with multile PUs which are satially or sectrally non-overlaing, as deicted in Fig. 1(a) and 1(b). Our aer makes the following contributions. First, we derive the otimal olicy for SU access in multile-pu systems with stationary general idle time distributions. We focus on a class of stationary access olicies, i.e. the same olicies are alied every time the channel becomes idle. We also look secifically at multile PU systems with exonential idle

2 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 2 time distributions, as well as idle time distributions that result in time-threshold otimal olicies. Finally, we extend our model to include two side information cases where extra information of PU activity is available to the SU. In the first case, the SU knows which PU is the last to become idle before an SU can transmit. In the second, the SU knows how long every PU has been idle before the channel becomes available. We investigate how such knowledge can affect the otimal olicy and erformance of the SU. The aer is organized as follows. In the next section, we resent some related work. We resent our system model in Section III and define the objective function in Section IV. In Section V we derive an otimal transmission scheme for a channel with multile PU constraints, where each PU has a general idle time distribution, and show that for certain idle time distributions, the otimal SU access olicy can be found in closed form. The system model is extended in Section VI such that the SU has side information on PU activity, and the otimal olicies with this new information are derived. In Section VII, we comare and analyze the erformance of our different olicies under the time caacity metric, and conclude our aer in Section VIII. II. RELATED WORKS In recent years, there has been an exlosion of research in cognitive radio. A large ortion of this research has been in sectrum overlay, where rotocols are devised to maximize SU sectrum utility when PUs are idle and rotect PU communication when they become busy. Within this aradigm, there are two focuses, satial and temoral domain research [3]. In the former, SU activity is assumed to occur in a much faster timescale than the PU activity, and hence the sectral environment (i.e. PU channel occuancy) is treated as static. The main roblem then becomes channel allocation between multile SUs given certain toologies, different channel availabilities, and interference between SUs. In [12], [13], [14], [15], the interference between SU nodes is modeled as a conflict grah, with varying methods and metrics used to allocate channels. In [16], [17], the authors formulate channel selection as a mixed integer linear rogramming roblem, under constraints on both ower and channel availability. In [18], the tradeoff between channel switching and maximizing SU bandwidth is considered under this assumtion. In all these works, PU activity is rotected since each SU node only has access to idle PU channels. Our work does not fall in this category since we assume that PUs change states in a timescale similar to that of SU activity. In the temoral domain category, PU activity varies quickly in the time domain and SUs within interference range must devise sensing and access schemes in concert to avoid significantly harming PU communication. As such, the metric used to measure interference to PUs is crucial. Several aers consider this ower, referred to as interference temerature, as the key metric. For examle, in [19], the authors consider multile SUs oerating in a multi-pu system where each PU has an average rate requirement and outage robability constraint, both functions of the interference ower caused by SUs. Power control for different states of PU activity is considered in [2]. However, in 27 the FCC officially ended consideration to establish an interference temerature standard for cognitive radio [21]. Their decision was based on several comments redicting a likely increase in interference with PUs in bands where it is used, stemming from the technical difficulties of imlementing such an aroach. Since we consider acket collision robability as our main constraint, we assume that any overla of SU and PU activity in the same band results in collision with the PU, which is a more conservative measure of PU interference. Like this work, other works have also considered acket collision robability. Several works have develoed medium access schemes for SUs under this rotection requirement [8], [9], [1]. One common formulation assumes a slotted system and is formulated using the Partially Observable Markov Decision Process (POMDP). For examle, [22], [23], consider a slotted structure network with a single PU rotection metric. Otimal access decisions are made by considering long observation history. Likewise, in [24], the authors consider a slotted system with a single CR, and model the roblem as a multi-armed bandit roblem to decide the best channel(s) to sense and access. In [11], the authors consider an overlay SU network on a multile PU network with slotted structure, where PU access deends on Markovian evolution. Our work is different from all of these because we consider multile different PU constraints that must be satisfied simultaneously. Many of these works also assume slotted activity, while our model is more general and can accomodate slotted and unslotted systems. Other works have considered unslotted systems as well. In [8], the authors introduce the erformance metric of time caacity, the average roortion of time that a SU can transmit without violating the PU s acket collision constraint, and generalize the results in [25]. The authors also extend their work to examine the imact of imerfect sensing on SU erformance as well as access issues between multile SUs in a single PU system. Our work in [26] generalizes the results for a single SU in [25] to the multile PU case for certain distributions, in articular for exonential distributions and distributions which result in time-threshold olicies. Our work in this aer resents a framework that can be used for general idle time distributions for multile PUs, and considers the effects of extra information about the PU system available to the SU. Much of the work in [25] can be generalized in this work as well. III. SYSTEM MODEL In this section we layout the model for PU and SU activity. Generally, we consider an SU that oerates within the interference range and on the same channels as multile noninterfering PU networks. Because the PUs are non-interfering, simultaneous transmission of multile PUs does not result in collision between them. A. Primary User Model In our system model, we assume there are M PUs that are indeendent and non-cooerative with the SUs and with

3 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 3 PU1 where PU2 V1 c i = lim Ni c, T N i PU V2 r and the PU rotection requirement is thus Sense Sense V =min(v1, V2 ) r c i η i, i 1,...,M}. (2) SU Collision w/ PU1 : Sense : Collision Collision w/ PU2 Collision w/ PU2 t We assume that the collision constraints for all PUs are known to the SU a riori. These constraints must be satisfied for all PUs. Fig. 2. Timing diagram for PU/SU system model. each other. A PU accesses its channel without sensing. PUs are non-interfering among themselves, i.e. they can transmit at the same time. This can be due to satial reuse as in Fig. 1(a), or due to communication on searate non-overlaing channels, as in 1(b). We assume acket-based transmission for all PUs in this aer. Fig. 2 illustrates the activity of each user. For PU i, V i denotes the idle time, which is governed by df/cdf f Vi ( )/F Vi ( ). Let E[V i ] = v i. The number of ackets transmitted er busy eriod is random and indeendent of PU i s idle times, and is denoted as N i, with E[N i ] = n i. We assume that the acket length is the same for all users, and normalize all activity to the length of a PU acket. Therefore, a busy eriod of the ith PU is N i. The robability that each PU is idle is then α i = vi v i+n i. All PU traffic is assumed to be stationary and ergodic. This is reasonable for acket-based data traffic, where the timescale for acket lengths is on the order of milliseconds but arrival rates are commonly stationary on the scale of hours [27]. We also define terms related to the union of the activity of all PUs in the system. We refer to this activity as unionized PU activity or simly the unionized PU. This activity rocess is also assumed to be stationary and ergodic. We denote the idle time of the unionized PU as V, which is governed by a robability distribution function f V ( ), with E[V ] = v. A busy eriod of the unionized PU is denotedn, with E[N ] = n Ṫhe robability that the unionized PU is idle is then α = v v +n = α i, (1) where the final equality results from the indeendence of the PU activity. The relationshi between each PU s activity and the unionized PU activity will be investigated in greater detail in Section IV. Each PU i also has a acket collision robability requirement denoted η i, defined as the maximum allowable robability of collision for a acket of the ith PU. Over a time interval [, T], we denote the number of ackets transmitted by the ith PU as N i, and the number of collisions exerienced by that user as Ni c. The collision robability of the ith PUs ackets exeriencing collision is denoted as c i = Pr[acket collision of ith PU], B. Secondary User Model Throughout this work we assume that there is a single SU, which may be a single CR radio or a SU basestation oerating within the range of multile PU networks [28]. We note that SUs may have access to multile bands, but low-cost SUs may have a set bandwidth requirement, and may not be able to frequently switch over channels at the timescales we are concerned with [25]. In [29], for examle, the authors state that channel switching times are often dwarfed by the minimum time of SU demand, and there is a risk of inefficiencies such as latency in switching channels too often. In that same work, a minimum allocatable bandwidth is a key factor in their model, which could lead to cases similar to the examle resented in Fig. 1(b) regardless of SU sensing and switching caabilities. Therefore, we assume that the SU is on a fixed channel and cannot transmit on a sub-band of the channel. We reiterate here that we are concerned with stationary otimal olicies. A stationary olicy is one that is alied whenever the channel becomes idle. We have: a) Packet Length: The SU slot length is denoted, and we assume v i, n i, and 1 (i.e. the length of a PU acket). In this aer, we study the extreme case where. In [8] it was roven that this results in the best SU caacity in the case of no overhead cost for a single PU channel, and overhead techniques used in that aer can be alied to our work as well. This assumtion simlifies analysis greatly, and our simulations show that non-zero acket lengths result in negligible differences. b) Sensing: We assume erfect sensing by the SU, i.e. that the SU can always detect the resence or absence of a PU, and that sensing time is negligible. The SU follows the listen-before-talk (LBT) rincile, where the SU senses the channel in each slot before allowing transmission. We assume that sensing occurs over the entire band of interest. The multile channel sensing roblem is a significant challenge and beyond the scoe of this work [28]. Our revious works consider imerfect sensing [8], [25], and these results aly here as well. We can also consider a fixed non-zero sensing time in the case of non-zero slot lengths. A fixed non-zero sensing time can reresent many commonly roosed sensing schemes, such as matched filter, energy, and feature detection. Since we assume a LBT scheme, the erformance in nonzero sensing cases is essentially a fixed fraction of the otimal erformance.

4 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 4 f Vi (t) F Vi (t) f r V (t) i FV r i (t) f V (t) F V (t) n i n α i α si f si(t) F si(t) h i(t) φ ij(t) TABLE I NOTATION FOR AN M-PU SYSTEM PDF of PU i idle time CDF of PU i idle time PDF of PU i residual idle time CDF of PU i residual idle time PDF of unionized PU idle time CDF of unionized PU idle time PU i average number of ackets er transmission average length of busy eriod for unionized PU Probability PU i idle Probability all PUs idle Probability PU i starts the idle eriod Idle time PDF given PU i starts the current idle eriod Idle time CDF given PU i starts the current idle eriod Portion of f V (t) due to PU i ending current idle eriod of length t Portion of f si(t) due to PU j ending current idle eriod of length t c) Collision Detection: Whenever an SU and PU transmit simultaneously, we assume that both exerience collision and the SU can detect the collision after the transmission. Perfect sensing, the SU acket length, and the LBT assumtion ensure that acket collisions occur only if a PU accesses the channel while an SU is already transmitting. This ensures that for any single PU busy eriod, at most one PU will exerience acket collision. Assuming erfect sensing and collision detection allows us to focus on investigating otimal caacity. Collisions with PUs are demonstrated in Fig. 2. d) Knowledge of Individual PUs: We assume that the SU has knowledge of the collision constraint, the idle time distribution, and the mean busy time of each individual PU a riori, i.e. f Vi ( ), n i, η i for all PUs, which are indexed by i. This knowledge can be obtained from the network usage histories obtained from network oerators [27]. This is feasible deending on the conditions of SU deloyment, where SU and PU network oerators may cooerate during initial SU network lanning and deloyment. For examle, in the DARPA rogram both PUs and SUs are military, so lanning of SU networks could conceivably include knowledge of PU network statistics [25]. e) Performance Metric: The SU s erformance metric is the time caacity, the ercentage of time that the SU can transmit successfully under the collision robability constraint. This metric is defined as below: C s = lim T SU s successful access time in [,T]. (3) T Since the channel observed by the SU has idle robability α, clearly C s α. We show through simulations later that a system with non-zero slot-length and sensing time results in a fraction of the otimal C s corresonding to the fraction of the slot time where transmission occurs. IV. OBJECTIVE FUNCTION The SU objective is to maximize its time caacity while satisfying all PU s collision constraints, max q Q subject to C s (q) c i (q) η i, i 1,...,M, (4) PU1 PU2 PU Fig. 3. V1 V2 r V =min(v1, V2) r r V1 V2 r V =min(v1, V2) Timing diagram showing two tyes of idle eriods. whereqis the set of all ossible transmission olicies that can be selected by the SU, while C s (q) and c i (q) are resectively the time caacity and robability of collision with user i given a olicy q( ) Q. Based on the system model, we now define the form of all olicies q( ) and derive the resentation of the objective function. The derivation assumes the system consists of a single SU co-existing with M non-interfering PUs. All notation is summarized in Table I. The objective function requires the derivation of the idle time distribution for the unionized PU activity. For brevity, we defer this derivation to Aendix A, and define the necessary terms here. Let t be the time elased since the beginning of the most recent idle eriod. We denotef si (t) as the idle time PDF of the unionized PU given that PUiis the last to sto transmission. In this case we say that PU i starts the idle eriod that follows. Then the unionized PU idle time distribution f V (t) can be rewritten as f V (t) = si f si (t), (5) where si is the robability that PU i starts an idle eriod. From Fig. 3, we can see that the residual idle time Vj r of PU j is also imortant to the system. The residual idle time distribution f V r j (t) is f V r j (t) = 1 F V j (t) = 1 F V j (t). (6) E[V j ] v j From the same figure, we observe that for M users, if PU i starts the idle eriod, V = min(v r 1,V r 2,...,V r i 1,V i,v r i+1,...,v r M ). We define φ ij (t) as φ ij (t) = f V r j (t)[1 F Vi (t)] [1 F V r k (t)], (7) k=1 k i,j which can be described as the ortion of the conditional PDF f si (t) due to PU j transmitting first after an idle eriod of length t. We note that f si (t) can now be defined as f si (t) = φ ij (t). (8) t

5 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 5 We also define h i (t) as h i (t) = sj φ ji (t), (9) which can be thought of as the ortion of the idle time PDF f V (t) that accounts for PU i transmitting first after an idle eriod of length t. We can rewrite f V (t) as f V (t) = h i (t). (1) We note here that, under the assumtion of ergodicity of the unionized PU activity, the equations for f V (t) can only be said to hold almost everywhere for an M PU system. Therefore, f V (t) and the resulting otimal olicy can only be said to aly with robability one instead of for all outcomes given a set of PU distributions. In all of the simulations in Section VII, all simulated instantiations of unionized PU activity and the resulting otimal olicies hold to the derivations made here, and we omit this fact for the sake of brevity in all following discussions. We now derive the objective function for our system. We define Φ (t) as the channel state of the unionized PU: Φ Idle, if all M PUs idle (t) = Busy, otherwise. We also define the general form of the SU olicy q(t) as the robability that the SU olicy transmits at time t, (t), if Φ (t) = Idle q(t) =, otherwise, where (t) 1. We note that this is a stationary olicy, where the same olicy is alied in every idle/busy cycle. Within each idle/busy cycle, the olicy is deendent on t, which reresents the time since the beginning of the most recent idle eriod. The structure of the otimal olicy q ( ) is the main focus of our work. For an SU olicy q( ), the time caacity equation (3) can now be defined as C s (q) = t f V (t) q(τ)dτdt v = G s(q) +n v, (11) +n where G s (q) = t f V (t) q(τ)dτdt. The numerator in Equation (11) calculates the average time transmitted in an idle/busy cycle of the unionized PU, while the denominator is the length of the average idle/busy cycle. The SU can collide with PU i only when that PU is the first to transmit following an idle eriod. Therefore, defining Ψ i (q) = q(t)h i (t)dt, from Equation (9) the constraint η i for PU i is satisfied under olicy q(t) if Ψ i (q) n i η i v +n v i +n i. (12) This equation can be understood as follows. Ψ i (q) is the robability that an SU using olicy q( ) will collide with a single acket from PU i in any given idle/busy cycle of the unionized PU with average duration v +n, as h i(t) is the ortion of the unionized PU idle time distribution function for when PU i is the first to begin transmitting after an idle eriod. At the same time, n i η i is the robability of collision with the SU that PU i can tolerate under its collision constraint for any given idle/busy cycle of PU i with average duration v i +n i. Therefore, a olicy q( ) must be determined that satisfies n i η i weighted by the ratio of these two different idle/busy cycle lengths. We now define η i = n v iη +n i v i+n i, and restate the objective function defined in (4) for the SU as max q(t): q(t) 1 C s (q) subject to Ψ i (q) η i, i 1,...,M}. (13) We note that the objective function (13) is in fact a convex otimization roblem. This is because the function C s (q) and the constraints Ψ i (q) from (13) are integral functions of q( ), where q(t) [,1] t, and integration is a linear function which thereby reserves convexity [3]. V. OPTIMAL POLICY FOR MULTIPLE PUS WITH GENERAL IDLE TIME DISTRIBUTIONS We now derive the otimal stationary SU access olicy in an M-PU environment with general idle time distributions. We first derive the most general form of the otimal olicy q ( ). We then resent a search algorithm for determining the otimal olicy based on rinciles of convex otimization. A. Deriving the General Otimal Policy We consider the following olicy q( ), defined as 1 F V (t) 1, if > M µ 1,Φ (t) = Idle ih i(t) 1 F q(t,µ) = V (t), if M = µihi(t) 1,Φ (t) = Idle, otherwise, (14) where µ = [µ 1,µ 2,...,µ M ] T. The condition for the iecewise boundaries of q(t, µ) resemble an inverted hazard function, where the df f V (t) is relaced by the sum of its h i (t) comonents weighted by µ. Each µ i term can be thought of as the imortance given to each PU s collision constraint. We then define the otimal olicy q (t) as q (t) = q(t,µ ), (15) where µ = [µ 1,µ 2,...,µ M ]T. Both µ and are chosen such that the following conditions are satisfied for i 1,...,M}: µ i, (C1) µ i = if q (t)h i (t)dt η i, q (t)h i (t)dt < η i. (C2) (C3)

6 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 6 Note here that µ i can be exlained as the Lagrange multilier for the ith constraint, andq ( ) is the function to be otimized. We now state the following theorem. Theorem 1: Policy q ( ) is an otimal olicy that maximizes the SU throughut while satisfying the collision constraints of all PUs. Proof: Consider any feasible olicy q( ) that satisfies the collision robability constraints of all PUs. We now rove that G s ( q) G s (q ), and therefore C s ( q) C s (q ). From (11), G s ( q) = f V (t) (a) G s ( q) = q(t) + µ i η i (b) + [ q (t) µ iη i = G s (q ) = G s (q ). t µ i q(τ)dτdt ( ) q(t)h i (t)dt η i 1 F V (t) [ 1 F V (t) µ i ] µ i h i(t) dt ] µ i h i(t) dt ( ) q (t)h i (t)dt η i The inequality (a) is true because q( ) must be a feasible olicy, which results in a collision robability less than or equal to the constraint. The inequality (b) results becauseq ( ) follows conditions (C1), (C2), and (C3), and q ( ) is ositive whenever 1 F V (t) M µ i h i(t) is ositive. We note that this olicy derivation is similar to that of the Lagrangian method [3]. However, it is by itself a roof, and does not make use of Lagrangian duality. Therefore, the otimal olicy search is in fact a search for vector µ such that the otimality conditions (C1), (C2), (C3) are satisfied. Requirement (C1) can be imosed by searching over only non-negative values for µ, but the others are deendent on q(t, µ). B. Otimal Policy for M PUs with Exonential Idle Times It is a common assumtion that channel idle time is distributed exonentially. For examle, in [27], the authors erform a measurement study that shows that the exonential call arrival model is adequate for cellular networks. It was also shown in [25] that the exonential case also rovides a lower bound to the achievable time caacity of any PU system. The same result, as we will show, alies here also. Therefore, in a system where the PUs average idle times are known but the idle time distributions are not, the olicy derived here can be used while guaranteeing the acket collision robability constraints of all users. When all M PUs have exonential idle time distributions, we can show that the otimal olicy q ( ) is q, ifφ (t) = Idle (t) = (16), otherwise, where [,1] is determined by the constraint conditions. In such a system the SU has equal robability of transmitting any time the channel is idle, so the time caacity isc s = α. This comes from the idle time distributions. Since all distributions are memoryless, f V r i (t) = f Vi (t), and F Vi (t) = F V r i (t) = 1 e λit for all i. Since V = min(v1 r,v 2 r,...,v i,...,vm r ), This also means that F V (t) = 1 e ( M λi)t. (17) h i (t) = C i e ( M λi)t, i 1,...,M}, (18) where C i is some constant. From (17) and (18), we observe that for any µ, there are three ossibilities: 1. 1 F V (t) < M µ ih i (t) for t F V (t) > M µ ih i (t) for t F V (t) = M µ ih i (t) for t. If the first is true, then from (15) we have q(t) = for all t, while if the second is true, q(t) = 1 for all t, imlying that the collision constraints are trivial. Therefore, assuming nontrivial collision constraints, an aroriate µ satisfies condition three, and (15) reduces to (16). Therefore, determining µ is not necessary, because we only need. We now determine a that satisfies (2) for all PUs. We note that can be found by alying (16) and (12), so we resent an intuitive exlanation here. Observing that is the robability that a single collision occurs during a single cycle of the unionized PU activity, this means a collision with PU i only occurs if the other M 1 PUs are idle when PU i begins transmitting. Therefore: j i α j = α α i n i η i, where the right side results from the fact that only one acket collision can occur when any PU becomes busy again, and on average PU i transmits n i ackets er busy eriod. It follows that can be written: = min (α i i 1...M} α n iη i,1). (19) Given this result for, we obtain the following theorem. Theorem 2. The maximum time caacity of the SU in a channel with the unionized activity of M PUs with exonential idle time distributions is: C s = min i 1...M} (C si,α ). (2)

7 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 7 PU1 PU2 V1 where Ti is determined by T i h i (τ)dτ = η i. (23) PU SU Fig. 4. PU1 Protects PU2 PU2 Protects PU1 Collision Protection with 2 PUs. t V2 r V =min(v1,v2 ) r PU2 Protects PU1 where C si = α i n i η i, which is the time caacity that can be achieved in the single-pu system consisting of the ith PU with exonential idle time. Proof: This results directly from (19) and (11). Another roof of this theorem is also available in our earlier work [26]. This result has the following intuitive exlanation. As Fig. 4 demonstrates, the PUs rotect each other from collision. This reduces the collision robability of each PU, allowing the SU to transmit more aggressively. This offsets the reduced channel availability. In some cases, all PUs will be rotected to the oint that their collision robability constraints cannot be violated, and the SU will transmit whenever the channel is idle, resulting in = 1 with time caacity C s (q ) = α. In addition, we also note that can be calculated simly using (19), which greatly reduces calculation time. This mutual rotection also benefits PU systems with general idle time distributions, but does not result in the same simlified structure as the exonential case. C. Time-Threshold Policy There are many cases in which the PU idle distributions result in a time-threshold olicy. For examle, systems where a PU has a uniform idle distribution or Weibull distributions with shae arameter under unity commonly result in a timethreshold olicy. Therefore, we investigate this case here. A time-threshold olicy begins transmitting as soon as the channel becomes idle for a length of time T, or until the channel becomes busy: q 1, if t T,Φ (t) = Idle (t) = (21), otherwise. We now state the following theorem. Theorem 3. If 1 F V (t)/h i (t) can be shown to be monotonically decreasing for i = 1,...,M, then the otimal olicy is a time-threshold olicy of the form in (21). Furthermore, T is T = min(t1,t 2,...,T M ), (22) t Proof: The condition for this olicy results directly from (15), which can be rewritten as 1 F V (t) 1, if > M q µ i hi(t) 1,Φ (t) = Idle (t) = 1 F V (t), if M µ = i hi(t) 1,Φ (t) = Idle, otherwise. Clearly, if the stated condition is satisfied, the above equation reduces to (21). Ti can be interreted as the otimal timethreshold for the SU if only PU i is constrained. Since only the minimum Ti can guarantee that all PU constraints are satisfied, the otimal olicy follows (21) with T as defined by (22) and (23). This leads to significant comutation reduction in determining the otimal SU olicy, since the olicy can now be found through (23). As stated before, several common distributions result in this tye of olicy. D. Otimal Policy Search Because the olicy search is essentially an otimization over the Lagrangian multilier vector µ, we can use convex otimization techniques to search for the otimal olicy. Therefore, in our numerical results, we use a descent method with line search to determine µ [3]. The idea of the olicy search is that a search or descent direction is chosen based on a current choice of µ, a distance to search in that descent direction is chosen based on a line search algorithm, and µ is udated. This rocess reiterates until a stoing criterion is reached. For ease of discussion, we define Ψ i (q) and L(q,µ) as Ψ i (µ) = L(µ) = G s (q(t,µ))+ q(t,µ)h i (t)dt, µ i (η i Ψ i(µ)). Ψ i (µ) is the collision robability achieved with PU i given olicy q(t, µ) from (12), and L(µ) is essentially the Lagrangian dual for the objective function (13). Beginning with an arbitrary, feasible choice for µ, the algorithm first determines a search direction µ using the gradient descent method, µ = [Ψ 1 (q(t,µ)) η 1,...,Ψ M (q(t,µ)) η M]. This direction is in fact the negative gradient of L(q,µ) with resect to µ. This direction increases the µ i corresonding to violated constraints under the current olicy, and decreases the µ i for PU constraints which are obeyed. This can be analogized as giving more/less weight to the stricter/looser PU constraints in the search. Next, a line search algorithm is used to determine the ste size ν, the distance to increment in the search direction. In

8 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 8 our numerical results we use a modified backtracking line algorithm [3]. It is essentially the same as the standard algorithm, but with conditions that revent the resulting ν from searching into negative µ i values which would violate condition (C1). Finally, µ is udated as µ µ+ν µ. If the udated µ satisifes conditions (C2) and (C3) to within a reasonable limit, the algorithm is terminated with the current iterate as µ. In ractice, this algorithm runs quickly. For all results in Section VII, each 2-PU-system olicy is found in under.3 seconds (and in many cases, significantly less) running on a Pentium GHz rocessor with 1 GB of RAM. Given that any olicy is meant to be oerative on the order of hours, this time is negligible. VI. PU SIDE INFORMATION We now investigate how extra information of PU activity can affect the otimal olicy, and ultimately the time caacity of the SU. In addition to the original system model assumtions, this information is assumed to be available to the SU through augmented sensing caabilities from the original system model. These two cases of extra PU information are as follows: Side Information 1: The SU knows which PU was the last to transmit. Side Information 2: The SU knows which PU was the last to transmit and how long the other PUs have been idle rior to the unionized channel going idle. We use SI-1 and SI-2 to describe these two cases, and No-SI to refer to the original olicy laid out in Section V. We note that in general the sensing caability of SI-1 would seem to imly the caability of SI-2, i.e. if a sensor can differentiate PUs, it is easy to assume it can also kee track of how long each has been idle. SI-1 and SI-2 can achieve significantly better erformance than the original case. Through our derivations we show that SI-2 is comutationally intractable. We now derive the new otimal olicy for each case. A. Otimal Policy for SI-1 We wish to leverage the new information when determining the otimal olicy q (t) of the SU. Naturally, the SU can act differently based on which PU ends transmission last. Therefore, an SU olicy q(t) can be written as q(t) = q i (t), if PU i is last to transmit, Φ (t) = Idle, otherwise. (24) where q i (t) is the SU access olicy when PU i is the last to transmit. From (12) and (13), the time caacity and the collision constraint can be rewritten as M C s (q) = si q i (t)[1 F si (t)]dt v +n, (25) sj q j (t)φ ji (t)dt min(η i,1). (26) where F si (t) is the cdf of the idle time distribution given that PU i is the last to transmit. Using these more secific equations, the objective function still follows the form of (13). We can see that the otimal olicy q (t) is similar to (14), excet that there are now M olicies for when each PU is the last to transmit. Therefore, we now define the otimal olicy qi (t,µ) given that PU i is the last to transmit for a given multilier vector µ: qi (t,µ) = 1, if 1 F si (t) > M µ jφ ij (t),φ (t) = Idle, if 1 F si (t) = M µ jφ ij (t),φ (t) = Idle, otherwise. (27) We note that this equation is quite similar to (15). F V (t) is relaced with F si (t), and h i (t) is relaced with φ ij (t), since the olicy being defined is only oerative when PU i is the last to transmit. We now define the otimal olicy q (t) as q (t) = q i (t,µ ), if PU i is last to transmit,φ (t) = Idle, otherwise, (28) where µ satisfies the conditions laid out in Section V-A. The roof of the otimality of the olicy is analogous to that of Theorem 1. This olicy can be obtained using a slightly modified version of the algorithm resented in Section V-D. B. Otimal Policy for SI-2 We now derive the otimal olicy for the SI-2 case. As stated reviously, this case is comutationally intractable and rimarily serves as an uer bound for ossible erformance for stationary olicies under our model. In the SI-2 case, the SU knows how long every PU has been idle rior to the channel going idle. We call this time the lost time, denoted by X i for the ith PU, with x i as a realization of X i. By definition, at least one PU s lost time will be zero at the start of any idle eriod, since at least one PU is transmitting rior to the moment the unionized PU goes idle. We define the vector X i as the vector containing all of the lost times given PU i is the last to return: X i = X 1,...,X i 1,,X i+1,...,x M } and x i = x 1,...,x i 1,,x i+1...x M } as a articular instantiation of X i. We note that the ith element is always zero for x i. The SU olicy deends on both the PUs idle time distributions and the vector X i. We denote the SU access olicy given that PU i is the last to transmit with lost time vector X i as q(t X i ), and a SU olicy q(t) as q(t) = q(t x i ), if X i =, X i = x i,φ (t) = Idle. (29) The time caacity and constraint equations (11) and (12) must be re-derived using the idle time distribution conditioned on the values of the lost time vector X i.

9 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 9 To obtain the time caacity and constraint equations, we must determine several related dfs conditioned on the lost time vector. We first define f si (t x i ), the conditional df of the idle time given that PU i is the last PU to transmit with lost time vector x i, f si (t x i ) = φ ij (t x i ). (3) In this case, φ ij (t x i ) is the ortion of f si (t x i ) due to PU j ending an idle eriod of length t, given that PU i is the last to transmit with a lost time vector of x i. We refer the reader to Aendix A for the derivation φ ij (t). If j = i we have If j i, we have: φ ii (t x i ) = f Vi (t) φ ij (t x i ) = f V j (t+x j ) 1 F Vj (x j ) j i k=1 k j 1 F Vj (t+x j ). 1 F Vj (x j ) (31) 1 F Vk (t+x k ). 1 F Vk (x k ) (32) We must also determine the distribution of the lost time vector X i. The distribution for X i is the same as the residual time distribution for the ith user: f Xi (x i ) = f V r i (x i ) = 1 F V i (x i ) v i, (33) and the distribution of X i is the joint distribution of all of the lost time distributions of the other variables. Since all PUs act indeendently, this distribution amounts to f Xi (x i ) = f Xj (x j ), (34) j i wheref Xi (x i ) is not included in (34) because inx i,pr[x i = ] = 1. With (3), (31), (32), and (34) we can derive the equation for f si (t) as f si (t) = f si (t x i )f Xi (x i )dx i. (35) x i The equation for f V (t) still follows (5). Integrating (3) over t also allows us to obtain the conditional cdf F si (t x i ). We also rewrite h i (t) in terms of φ ij (t x i ) as h i (t) = sj φ ji (t x i )f Xi (x i )dx i. (36) x i We are now able to derive the time caacity C s (q) and constraint equations for SI-2: q(τ)[1 F V C s (q) = (τ)]dτ v +n M = si x i q(τ x i )[1 F si (τ x i )]dx i dτ v +n, (37) sj x j q(t x j )φ ji (t x j )f Xj (x j )dx j dt η i. (38) The objective function then follows the form of (13) using the time caacity and constraint equations (37) and 38). As in the revious cases, we define the otimal olicy q (t) as the olicy that corresonds to the vector µ such that the otimality conditions (C1)-(C3) are satisfied. The otimal olicy q (t x i ) for PU i given x i is: q (t x i ) = 1, if 1 F si (t x i ) > M µ j φ ij(t x i ),Φ (t) = Idle, if 1 F si (t x i ) = M µ j φ ij(t x i ),Φ (t) = Idle, otherwise, (39) and the otimal olicy q (t) is q q (t x i ), if X i = and X i = x i, Φ (t) = Idle (t) =, otherwise, (4) where µ = µ 1,...,µ M } satisfies (C1)-(C3). The roof is the same as that resented in theorem 1. This olicy gives better erformance than either the No-SI or SI-1 cases because of the additional information. However, in ractice, obtaining the otimal olicy in (4) is comutationally difficult. For any µ, the collision robability deends on an infinite number of olicies q(t x i ) corresonding to all ossible lost time vectors, rendering numerical solutions rohibitively exensive. VII. NUMERICAL RESULTS We now resent numerical results demonstrating how SU erformance is affected by the number of PUs in the system and the side sensing information defined in the revious section. The results resented are meant to demonstrate the effect of time caacity in a number of different network scenarios where multile PUs may need to be rotected. Therefore, we leave out secific details of PHY/MAC layer network scenario assumtions. First, we investigate how the number of PUs affects the SU erformance for the No-SI case. We show that while time caacity generally degrades with the number of users, throughut increases to a oint if the PUs are on different channels. Then we show that in the SI-1 and SI-2 cases the extra PU information available to the SU over the No-SI system model leads to higher time caacity while still satisfying PU acket collision robability constraints. A. Multile PUs We first investigate how multile PUs affect SU erformance in the No-SI system model. In all simulations in this section, we have multile PUs with the same idle/busy time distributions,v i = 8 andn i = 2, and the same collision constraint η. We comare SU erformance for two different idle time distributions, exonential and uniform. The PU acket length is 1, and the SU acket length =.1. Therefore, any SU olicy q( ) must assign transmission robabilities for time

10 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 1 C s C s vs. Multile Channels, v i = 8, n i = 2, α =.8 Uniform Idle, η=.1 Exonential Idle, η=.1 Uniform Idle, η=.2 Exonential Idle, η=.2 Uniform Idle, η=.3 Exonential Idle, η=.3 α Fig. 6. Throughut vs. Number of PU Channels (M) with v i = 8,n i = 2. R s vs. M PU Channels, v i =8, n i =2, η=.1 R s P/N =1, unif P/N =1, ex P/N =1, unif P/N =1, ex Number of Channels Fig. 5. C s vs. Multile PUs, with single PU arameters v i = 8,n i = 2. increments of length, and all idle time distribution functions are calculated with time increment of also. First, in Fig. 5, we study the time caacity erformance as a function of the number of PUs to be rotected. This is an accurate measure of SU erformance in the satial searation case demonstrated in Fig. 1(a), since all PUs and the SU oerate on the same channel. In Fig. 5, each curve reresents a simulation where all PUs have the same rotection requirement and idle time distribution, either exonential or uniform. We first observe that for all cases, the time caacity is eventually limited by the idle robability α. This corresonds to when the PUs are sufficiently rotected such that no PU constraint can be violated. In this case, the SU transmits whenever the channel is idle. We also note that the time caacity for uniform idle distribution cases degrades as the number of PUs grows, with the largest dro between M = 1 and M = 2. However, the exonential idle time cases exerience no degradation until the SU transmits with robability 1, corresonding to a time caacity of α. Finally, we observe that the uniform case outerforms the exonential case for all η and M before the time caacity is limited by α. This is because the memorylessness of the exonential case reduces the SU access olicy to a random access scheme with robability corresonding to the collision constraint. On the other hand, in the uniform distribution case, the time that the unionized channel has been idle hels in redicting when the channel will become busy. This results in a olicy that exloits this redictability, transmitting more aggressively earlier on in an idle eriod. In a case where each PU resides on a different channel as in Fig. 1(b), we must also consider the bandwidth of the unionized channel to evaluate SU erformance. We consider this in Fig. 6. In these simulations, we now assume that in addition to each PU having the same usage statistics and collision constraints as in Fig. 5, each PU occuies its own channel of bandwidth B. All channels have the same noise ower densityn, and the SU oerates only under a maximum ower constraint P. We now define the throughut of the SU R s (M) as the roduct of the time caacity and the Shannon R s R s M (a) η =.1 R s vs. M PU Channels, v i =8, n i =2, η=.2 P/N =1, unif P/N =1, ex P/N =1, unif P/N =1, ex M (b) η =.2 R s vs. M PU channels, v i =8, n i =2, P/N =1 η=.1, unif η=.1, ex η=.2, unif η=.2, ex η=.3, unif η=.3, ex M (c) η varying, P/N = 1 caacity of an SU oerating on M channels. Therefore, ) P R s (M) = C s (M)MB log 2 (1+, (41) MB N where C s (M) is the time caacity of the SU on M channels. Therefore, for any simulation, C s (M) always follows one of the curves in 5. For simlicity, we also assume that B = 1. In Fig. 6(a) and 6(b), η is held constant at.1 and.2 resectively, and each curve reresents a different P/N ratio and idle time distribution (exonential vs. uniform). First, observing Fig. 5, we see that when η =.1 and.2, the number of PUs that results in a time caacity C s α is 8 and 5, resectively. Now, from 6(a) and 6(b), we see that when η =.1 and.2, the maximum throughut also occurs at 8

11 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 11 and 5 PU channels resectively, regardless of distribution and SU transmission ower. This suggests that a good heuristic to obtain maximum throughut in a multi-pu multi-channel system is to use as many channels as ossible, such that the time caacity is not severely limited by the idle robability of the channel. Fig. 6(a) and 6(b) also show that the uniform case has a higher throughut than its corresonding exonential case until the maximum throughut is reached. This is again due to the memorylessness of the exonential case, which makes the arrival of the PU unredictable. However, we also notice that the exonential and uniform cases with the same SU ower and η value both result in roughly the same erformance at the elbow of each of the curves. The cause of this is obvious: in either case, the SU is transmitting whenever the channel is idle, and both cases have the same idle robabilities. This result suggests that a strong heuristic to obtaining a channel grouing that obtains near otimal erformance may be to treat all channels as exonential regardless of their actual distributions. This will significantly reduce comutation of the unionized PU idle time distribution as well as the otimal access olicy. Finally, in Fig. 6(c), P/N = 1 for all simulations, and each curve reresents a different η value and idle time distribution. We note that regardless of the η value, as the number of PU channels goes u, eventually erformance is limited by the time caacity. B. Comarison of No-SI and SI-1 We now comare the erformance of the No-SI and SI- 1 cases. In these simulations, activity of a two PU system is generated. Both PUs have uniform idle time distributions with v 1 = v 2 = 8, and general busy time distributions with n 1 = 2, n 2 = 4 ackets er transmission. Several busy time distributions were considered, including uniform, exonential, and deterministic acket length distributions, which all resulted in very similar results. We assume that the ackets for both PUs have the same length, so that α 1 =.8, α 2 =.6667, and α = PU and SU acket lengths are 1 and.1, resectively. Results with larger SU acket lengths, including a case where SU and PU acket lengths are equal, showed negligible differences in erformance with the cases resented here, and were therefore omitted. In Fig. 7(a), and SI-1 olicies are lotted as a function of η, where both PUs are assumed to have the same acket collision robability constraint, i.e. η 1 = η 2 = η. Table II(a) dislays the collision robability values obtained by the No-SI (NS) and SI-1 (S1) olicies for each PU and desired η value. In Tables II(a), II(b), and III, the constraint values set by the PUs are labeled as η i, and the collision robabilities achieved by the otimal olicies are labeled as ˆη i. We observe that the SI-1 otimal olicy achieves higher C s when η <.35. We also notice that, while both olicies are able to achieve the desired η 1 value, the SI-1 olicy is able to achieve higher collision rates than No-SI for PU2 without violating the collision constraints. This is because SI-1 exloits the extra information to transmit more aggressively, achieving a higher time caacity. Finally, at η =.35, we notice that both olicies achieve aroximately the same C s, and that both olicies have collision robability rates less than the desired η. This corresonds to the case where the SU transmits whenever the channel is idle: C s = α =.5333, and neither collision constraint can be violated. In Fig. 7(b), we observe the effect that differing collision constraint values have on the time caacity for No-SI and SI-1. Time caacity C s is lotted against changing η 2, while PU1 s collision constraint is held at η 1 =.1, with corresonding achieved collision robability values dislayed in Table II(b). Again it is clear that the SI-1 olicy erforms better in general, and similarly, SI-1 achieves a higher collision robability than No-SI without violating the collision constraints. The one excetion occurs at η 2 =.6, where both cases satisfy the collision constraints with equality (bolded in Table II(b)). The corresonding C s values achieved by No-SI and SI-1 are the same at this oint. In general, this observation holds true for all combinations of idle time distributions that result in T olicies: for any η 1 value, there is a corresonding η 2 value such that both constraints are satisfied with equality by the No-SI and SI-1 otimal olicies. In this case, the No- SI and SI-1 olicies are both time-threshold olicies with the same T value, which results in their corresondingc s values being equal. We also observe that both C s curves become constant at higher η 2 values. This is because η 1 becomes the limiting constraint for both olicies, such that neither No-SI nor SI-1 can transmit more aggressively to exloit looserη 2 constraints. Finally, in Table III, we observe the effect of non-zero SU slot lengths and sensing times on erformance. We run all olicies over the same simulated PU transmission atterns as in the results from Fig. 7, with v 1 = v 2 = 8, n 1 = 2, and n 2 = 4. PU slot length is again 1. We assume that the SU senses the channel for a constant amount of time in every slot, with s denoting the sensing time. The table in this case corresonds only to SI-1 olicies; the results from the No-SI case show similar trends. From the table, we see that erformance is affected very little by the increased acket length. In all cases with zero sensing time, when comared with the theoretical olicy erformance, there are only slight variations of achieved collision robabilities and time caacity. Generally, higher values of ˆη i corresond to slightly higher values of time caacity. These differences are minor, since the differences in ˆη i are less than 2% in all cases with resect to the =.1 theoretical case. In the cases with non-zero sensing time, we see that while achieved collision robabilities remain the same, C s decreases due to the time sent sensing the channel. The decreases in C s corresond to the s / factor sent for sensing in each time slot. Therefore, we conclude that for non-zero SU slot lengths, erformance remains within accetable limits assuming that the SU slot length is not greater than the PU acket length. C. SI-2 Policy Performance We now comare SI-2 erformance to the other two cases. Simulation arameters are the same as the revious section.

12 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 12 Fig. 7. Time caacity erformance for 2 PUs with uniform idle time distributions, v 1 = v 2 = 8, n 1 = 2,n 2 = 4. Time Caacity vs. η, η 1 =η 2 =η Time Caacity (C s ) vs. η 2, η 1 = C s.3.2 No SI SI 1 C s.15.1 No SI SI η 1 (a) Time Caacity vs. η = η 1 = η η 2 (b) Time Caacity vs. η 2 with η 1 =.1 TABLE II ACHIEVEDη 1,η 2,C s FOR FIG. 7 (a) Achieved values for Fig. 7(a), η = η 1 = η 2 η ˆη 1 (NS) ˆη 2 (NS) ˆη 1 (S1) ˆη 2 (S1) (b) Achieved values for Fig. 7(b), η 2 varying, η 1 =.1 η 2 ˆη 1 (NS) ηˆ 2 (NS) ˆη 1 (S1) ˆη 2 (S1) TABLE III SU PERFORMANCE FOR SI-1 POLICY WITH NONZERO SLOT LENGTH, SENSING TIME s η 1 =.1,η 2 =.6 η 1 =.1,η 2 =.1 Policy Tye C s ηˆ 1 ˆη 2 C s ηˆ 1 ηˆ 2 =.1, s = (Theoretical) =.1, s = = 1, s = = 1, s = = 1, s = For SI-2 simulations, otimal search is imractical because of heavy comutation. Instead, we use a randomly selected (µ 1,µ 2 ) air to generate an arbitrary, simlified SI-2 olicy, i.e. no olicy search is initiated. Since the SI-2 olicy is deendent on the lost time vector X i, for any (µ 1,µ 2 ) air there are an infinite number of olicies corresonding to different lost times, meaning that the achieved collision robability of any SI-2 olicy is comutationally exensive. To make this more tractable, we divide the ossible lost time into intervals of width W. For each interval, a olicy is determined in a manner similar to (39). For a 2-PU system with a given (µ 1,µ 2 ) air, q(t x i, x j W = k) = 1 F 1, if si(t x j=kw) > 1, M µ Φ (t) = Idle jφ ij(t x j=kw) 1 F, if si(t x j=kw) M µ jφ ij(t x j=kw) = 1, Φ (t) = Idle, otherwise, (42) where x i =, and x j is the lost time of PU j with j i. The modified olicy is then q(t x i, xj q(t) = W = k), if X i = x i, Φ (t) = Idle, otherwise. (43) Although the olicy defined in (43) reduces comutation greatly, it is still exensive to search for an otimal(µ 1,µ 2 ) for a given η 1,η 2 air. Instead, we obtain an arbitrary SI-2 olicy using a randomly selected (µ 1,µ 2 ) air. We then comare the erformance of this olicy to otimal No-SI and SI-1 olicies. Using the collision robabilities obtained under the SI-2 olicy, we erform the otimal olicy search to determine the No-SI and SI-1 otimal olicies corresonding to those collision robabilities. Table IV shows the results of one such comarison. We use a lost time interval width W = 1, and obtain a SI-2 olicy that yields η 1 =.11,η 2 =.11. The otimal search algorithm is run for No-SI and SI-1 using those constraints. The SI-2 olicy achieves a 9.2% gain over SI-1, and a 34.3%

13 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 13 TABLE IV PERFORMANCE COMPARISON BETWEEN 3 CASES, WITHη 1 =.11, η 2 =.11, AND LOST INTERVAL WIDTHW = 1. gain over the No-SI. Policy η 1 achieved η 2 achieved C s No-SI SI SI VIII. CONCLUSION In this aer, we studied oortunistic sectrum access in a system with multile PUs, where the PUs are heterogeneous in terms of idle time distribution and acket collision robability requirement. We determined the form of the otimal olicy, and secified two secial cases for which the otimal olicy is easier to determine: the case where all PUs have exonential idle times, and the case where the PU idle time distributions lead to a time-threshold olicy. We then studied the effect that extra information of PU activity has on the SU otimal olicy. The first case studied was one in which the SU knows which PU was the last to transmit before the channel becomes idle. In the second, the SU had the additional knowledge of how long each PU had been idle before all PUs became idle. For both cases, otimal olicies were also determined. Simulations results demonstrated that the extra PU information significantly increases SU erformance. These results have imlications for future work in channel ooling and allocation in SU networks. One such roblem is how an SU should otimally access a set of PU channels because there is a tradeoff between instantaneous data rate and time caacity. Another roblem is channel allocation for multile SUs. In general, revious work on allocation has assumed a static sectral environment, i.e. licensed channels are idle or busy for long eriods of time. This work imlies new allocation scenarios in which PUs have frequent idle/busy transitions, and SUs must consider PU activity statistics when allocating channels since certain combinations of channels may be better groued together. APPENDIX A DERIVATION OF UNIONIZED PU IDLE TIME DISTRIBUTION From (5), we need to derive si andf si (t). To determine si, the activity of the PUs is observed over time interval [,T], where T is the length of the first N idle/busy cycles of the unionized PU. Ik i,bi k,i k, and B k are the kth idle and busy eriods of the unionized PU and PU i resectively. Therefore, T = I k +B k N k=1 N i = ( Im i +Bi m )+R i, i 1,...,M}, m=1 (A-1) where R i < I 1 N i+1 +B1 N i+1, and in general N i N j N for i,j 1,...,M. Since PU i must end a transmission for every idle/busy cycle, each PU ends N i transmissions during T. An idle eriod started by PU i results if all other PUs are idle when PU i transits from busy to idle state. Defining N si as the number of idle eriods started by PU i, we know that as T, with lim T N si = N i lim T α j, j i N si = N. With these equations, si can be shown to be si = 1 v i M 1 v j. (A-2) We now determine f si (t). From (8), we need to derive φ ij (t). If i = j, then φ ij (t) = φ ii (t) is φ ii (t) = Pr[min(V1 r,v 2 r,...,v i,...,vm r ) = V i] = f Vi (t) [1 F V r j (t)]. j i (A-3) If i j, φ ij (t) is the robability that the residual idle time of PU j is t, and that all other users idle times are greater than t, or φ ij (t) = Pr[min(V1 r,v2 r,...,vj r,...,v i,...,vm) r = Vj r ] = f V r j (t)[1 F Vi (t)] [1 F V r k (t)]. k=1 k i,j (A-4) We have now derived all terms from (5) for the full unionized idle time distribution f V (t). From (5) we can obtain v, and combining with (1), we can obtain n as n = v ( 1 α 1). REFERENCES (A-5) [1] Sectrum olicy task force reort, Federal Communications Commision, Reort, Et docket No , November 22. [2] Facilitating oortunities for flexible, efficient, and reliable sectrum use emloying cognitive radio technologies, notice of roosed rule making and order, Federal Communications Commision, Reort, Et docket No , December 23. [3] Q. Zhao and B. Sadler, A survey of dynamic sectrum access: Signal rocessing, networking, and regulatory olicy, IEEE Signal Processing Magazine, vol. 55, no. 5, , 27. [4] S. Haykin, Cognitive radio: brain-emowered wireless communications, IEEE Journal on Selected Areas in Communications, vol. 23, no. 2, , 25, [5] J. Mitola and J. Maguire, G. Q., Cognitive radio: making software radios more ersonal, Personal Communications, IEEE [see also IEEE Wireless Communications], vol. 6, no. 4, , 1999, [6] Shared sectrum comany: DARPA XG rogram information. [Online]. Available: htt:// [7] F. W. Seelig, A descrition of the august 26 xg demonstrations at fort a.. hill, in Second IEEE International Symosium on Dynamic Sectrum Access Networks, DySPAN, 27, [8] S. Huang, X. Liu, and Z. Ding, Oortunistic sectrum access in cognitive radio networks, IEEE INFOCOM 28, Phoenix, AZ, USA, Aril, 28. [Online]. Available: htt:// senhua/osainfocom8.df

14 JUNG AND LIU: OPPORTUNISTIC SPECTRUM ACCESS IN MULTIPLE PRIMARY USER ENVIRONMENTS UNDER THE PACKET COLLISION CONSTRAINT 14 [9] Q. Zhao, S. Geirhofer, L. Tong, and B. M. Sadler, Otimal dynamic sectrum access via eriodic channel sensing, in Proc. Wireless Communications and Networking Conference (WCNC), 27. [1] T. Shu, S. Cui, and M. Krunz, Medium access control for multi-channel arallel transmission in cognitive radio networks, Proceedings of the IEEE GLOBECOM 26 Conference, San Francisco, CA, Dec., 26. [11] R. Urgaonkar and M. J. Neely, Oortunistic scheduling with reliability guarantees in cognitive radio networks, IEEE INFOCOM 28, Phoenix, AZ, USA, Ar., 28. [12] W. Wang and X. Liu, List-coloring based channel allocation for oensectrum wireless networks, Proceedings of the IEEE Vehicular Tech. Conference, , 25. [13] H. Zheng and C. Peng, Collaboration and fairness in oortunistic sectrum access, Proceedings of the IEEE ICC Conference, 25. [14] L. Cao and H. Zheng, Understanding the ower of distributed coordination for dynamic sectrum management, ACM/Sringer Mobile Networks and Alications (MONET),, vol. 13, [15] L. Yang, L. Cao, and H. Zheng, Physical interference driven dynamic sectrum management, Proceedings of IEEE Symosium on New Frontiers in Dynamic Sectrum Access Networks(DySPAN). [16] Y. S. Y. Thomas Hou and H. D. Sherali, Sectrum sharing for multiho networking with cognitive radios, IEEE Journal on Selected Areas in Communications, vol. 26, no. 1, 28. [17] T. Shu and M. Krunz, Coordinated channel access in cognitive radio networks: A multi-level sectrum oortunity ersective, IEEE INFO- COM 29, Rio De Janeiro, Brazil, Br, 29. [18] D. Xu, E. Jung, and X. Liu, Otimal bandwidth selection in multichannel cognitive radio networks: How much is too much? Third IEEE International Symosium on Dynamic Sectrum Access Networks, DySPAN, 28. [19] D. I. Kim, L. B. Le, and E. Hoossain, Joint rate and ower allocation for cognitive radios in dynamic sectrum access environment, IEEE Transactions on Wireless Communications, vol. 7, no. 12, , 28. [2] Y. Chen, G. Yu, Z. Zhang, H.-H. Chen, and P. Qiu, On cognitive radio networks with oortunistic ower control strategies in fading channels, IEEE Transactions on Wireless Communications, vol. 7, no. 7, , 28. [21] In the matter of establishment of an interference temerature metric to quantify and manage interference and to exand available unlicensed oeration in certain fixed, mobile and satellite frequency bands, Federal Communications Commision/Office of Engineering and Technology, OET Reort, FCC 7-78, May 27. [22] Q. Zhao, L. Tong, A. Swami, and Y. Chen, Decentralized cognitive mac for oortunistic sectrum access in ad hoc networks: A omd framework, IEEE Journal on Selected Areas in Communications, vol. 25, no. 3, , 27. [23] Y. Chen, Q. Zhao, and A. Swami, Joint design and searation rincile for oortunistic sectrum access in the resence of sensing errors, Proc. of IEEE Asilomar Conference on Signals, Systems, and Comuters, November,, 26. [24] T. Javidi, B. Krishnamachari, Q. Zhao, and M. Liu, Otimality of myoic sensing in multi-channel oortunistic access, Proc. of IEEE International Conference on Communications (ICC), May, 28. [25] S. Huang, X. Liu, and Z. Ding, Otimal transmission strategies for dynamic sectrum access in cognitive radio networks, IEEE Transactions on Mobile Communications, vol. 8, no. 12, 29. [26] E. Jung and X. Liu, Oortunistic sectrum access in heterogeneous user environments, Third IEEE International Symosium on Dynamic Sectrum Access Networks, DySPAN, 28. [27] D. Willkomm, S. Machiraju, J. Bolot, and A. Wolisz, Primary users in cellular networks: A large-scale measurement study, Proceedings of IEEE Symosium on New Frontiers in Dynamic Sectrum Access Networks(DySPAN). [28] I. F. Akyildiz, W.-Y. Lee, M. C. Vuran, and S. Mohanty, A survey on sectrum management in cognitive radio networks, IEEE Communications Magazine, Aril 28. [29] Y. Yuan, P. Bahl, R. Chandra, T. Moscibroda, and Y. Wu, Allocating dynamic time-sectrum blocks in cognitive radio networks, in MobiHoc 7: Proceedings of the 8th ACM international symosium on Mobile ad hoc networking and comuting. New York, NY, USA: ACM, 27, [3] S. Boyd and L. Vandenberghe, Convex Otimization, 1st ed. New York, NY, U.S.A.: Cambridge University Press, 24. Eric Jung is currently a Ph.D. candidate in the Electrical and Comuter Engineering Deartment at University of California, Davis. He received his B.S. and M.S. in electrical engineering from the same university in 26 and 29, resectively. His research interests include wireless networking and communications, with an emhasis on dynamic sectrum access. Xin Liu received the PhD degree in electrical engineering from Purdue University in 22. She is currently an associate rofessor in the Comuter Science Deartment at the University of California, Davis. Before joining UC Davis, she was a ostdoctoral research associate in the Coordinated Science Laboratory at the University of Illinois at Urbana- Chamaign. Her research is on wireless communication networks with a focus on resource allocation and dynamic sectrum management. She received the Best Paer of Year Award from the Comuter Networks Journal in 23 for her work on oortunistic scheduling. She received the US National Science Foundation CAREER Award in 25 for her research on Smart-Radio- Technology-Enabled Oortunistic Sectrum Utilization. She received the Outstanding Engineering Junior Faculty Award from the College of Engineering, UC Davis, in 25. She is a member of the IEEE.

Optimal Random Access and Random Spectrum Sensing for an Energy Harvesting Cognitive Radio with and without Primary Feedback Leveraging

Optimal Random Access and Random Spectrum Sensing for an Energy Harvesting Cognitive Radio with and without Primary Feedback Leveraging 1 Otimal Random Access and Random Sectrum Sensing for an Energy Harvesting Cognitive Radio with and without Primary Feedback Leveraging Ahmed El Shafie, Member, IEEE, arxiv:1401.0340v3 [cs.it] 27 Ar 2014

More information

Delay Performance of Threshold Policies for Dynamic Spectrum Access

Delay Performance of Threshold Policies for Dynamic Spectrum Access IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL., NO. 7, JULY 83 Delay Performance of Threshold Policies for Dynamic Sectrum Access Rong-Rong Chen and Xin Liu Abstract In this aer, we analyze the delay

More information

MATHEMATICAL MODELLING OF THE WIRELESS COMMUNICATION NETWORK

MATHEMATICAL MODELLING OF THE WIRELESS COMMUNICATION NETWORK Comuter Modelling and ew Technologies, 5, Vol.9, o., 3-39 Transort and Telecommunication Institute, Lomonosov, LV-9, Riga, Latvia MATHEMATICAL MODELLIG OF THE WIRELESS COMMUICATIO ETWORK M. KOPEETSK Deartment

More information

Convex Optimization methods for Computing Channel Capacity

Convex Optimization methods for Computing Channel Capacity Convex Otimization methods for Comuting Channel Caacity Abhishek Sinha Laboratory for Information and Decision Systems (LIDS), MIT sinhaa@mit.edu May 15, 2014 We consider a classical comutational roblem

More information

Combining Logistic Regression with Kriging for Mapping the Risk of Occurrence of Unexploded Ordnance (UXO)

Combining Logistic Regression with Kriging for Mapping the Risk of Occurrence of Unexploded Ordnance (UXO) Combining Logistic Regression with Kriging for Maing the Risk of Occurrence of Unexloded Ordnance (UXO) H. Saito (), P. Goovaerts (), S. A. McKenna (2) Environmental and Water Resources Engineering, Deartment

More information

Optimal Random Access and Random Spectrum Sensing for an Energy Harvesting Cognitive Radio with and without Primary Feedback Leveraging

Optimal Random Access and Random Spectrum Sensing for an Energy Harvesting Cognitive Radio with and without Primary Feedback Leveraging Otimal Random Access and Random Sectrum Sensing for an Energy Harvesting Cognitive Radio with and without Primary Feedback Leveraging Ahmed El Shafie Wireless Intelligent Networks Center WINC, Nile University,

More information

MODELING THE RELIABILITY OF C4ISR SYSTEMS HARDWARE/SOFTWARE COMPONENTS USING AN IMPROVED MARKOV MODEL

MODELING THE RELIABILITY OF C4ISR SYSTEMS HARDWARE/SOFTWARE COMPONENTS USING AN IMPROVED MARKOV MODEL Technical Sciences and Alied Mathematics MODELING THE RELIABILITY OF CISR SYSTEMS HARDWARE/SOFTWARE COMPONENTS USING AN IMPROVED MARKOV MODEL Cezar VASILESCU Regional Deartment of Defense Resources Management

More information

Approximate Dynamic Programming for Dynamic Capacity Allocation with Multiple Priority Levels

Approximate Dynamic Programming for Dynamic Capacity Allocation with Multiple Priority Levels Aroximate Dynamic Programming for Dynamic Caacity Allocation with Multile Priority Levels Alexander Erdelyi School of Oerations Research and Information Engineering, Cornell University, Ithaca, NY 14853,

More information

On the Role of Finite Queues in Cooperative Cognitive Radio Networks with Energy Harvesting

On the Role of Finite Queues in Cooperative Cognitive Radio Networks with Energy Harvesting On the Role of Finite Queues in Cooerative Cognitive Radio Networks with Energy Harvesting Mohamed A. Abd-Elmagid, Tamer Elatt, and Karim G. Seddik Wireless Intelligent Networks Center (WINC), Nile University,

More information

Multi-Channel MAC Protocol for Full-Duplex Cognitive Radio Networks with Optimized Access Control and Load Balancing

Multi-Channel MAC Protocol for Full-Duplex Cognitive Radio Networks with Optimized Access Control and Load Balancing Multi-Channel MAC Protocol for Full-Dulex Cognitive Radio etworks with Otimized Access Control and Load Balancing Le Thanh Tan and Long Bao Le arxiv:.3v [cs.it] Feb Abstract In this aer, we roose a multi-channel

More information

An Analysis of TCP over Random Access Satellite Links

An Analysis of TCP over Random Access Satellite Links An Analysis of over Random Access Satellite Links Chunmei Liu and Eytan Modiano Massachusetts Institute of Technology Cambridge, MA 0239 Email: mayliu, modiano@mit.edu Abstract This aer analyzes the erformance

More information

s v 0 q 0 v 1 q 1 v 2 (q 2) v 3 q 3 v 4

s v 0 q 0 v 1 q 1 v 2 (q 2) v 3 q 3 v 4 Discrete Adative Transmission for Fading Channels Lang Lin Λ, Roy D. Yates, Predrag Sasojevic WINLAB, Rutgers University 7 Brett Rd., NJ- fllin, ryates, sasojevg@winlab.rutgers.edu Abstract In this work

More information

Information collection on a graph

Information collection on a graph Information collection on a grah Ilya O. Ryzhov Warren Powell February 10, 2010 Abstract We derive a knowledge gradient olicy for an otimal learning roblem on a grah, in which we use sequential measurements

More information

Feedback-error control

Feedback-error control Chater 4 Feedback-error control 4.1 Introduction This chater exlains the feedback-error (FBE) control scheme originally described by Kawato [, 87, 8]. FBE is a widely used neural network based controller

More information

Improved Capacity Bounds for the Binary Energy Harvesting Channel

Improved Capacity Bounds for the Binary Energy Harvesting Channel Imroved Caacity Bounds for the Binary Energy Harvesting Channel Kaya Tutuncuoglu 1, Omur Ozel 2, Aylin Yener 1, and Sennur Ulukus 2 1 Deartment of Electrical Engineering, The Pennsylvania State University,

More information

Optimal Power Control over Fading Cognitive Radio Channels by Exploiting Primary User CSI

Optimal Power Control over Fading Cognitive Radio Channels by Exploiting Primary User CSI Otimal Power Control over Fading Cognitive Radio Channels by Exloiting Primary User CSI Rui Zhang Abstract arxiv:0804.67v2 [cs.it] 7 Feb 2009 This aer is concerned with sectrum sharing cognitive radio

More information

Age of Information: Whittle Index for Scheduling Stochastic Arrivals

Age of Information: Whittle Index for Scheduling Stochastic Arrivals Age of Information: Whittle Index for Scheduling Stochastic Arrivals Yu-Pin Hsu Deartment of Communication Engineering National Taiei University yuinhsu@mail.ntu.edu.tw arxiv:80.03422v2 [math.oc] 7 Ar

More information

Towards understanding the Lorenz curve using the Uniform distribution. Chris J. Stephens. Newcastle City Council, Newcastle upon Tyne, UK

Towards understanding the Lorenz curve using the Uniform distribution. Chris J. Stephens. Newcastle City Council, Newcastle upon Tyne, UK Towards understanding the Lorenz curve using the Uniform distribution Chris J. Stehens Newcastle City Council, Newcastle uon Tyne, UK (For the Gini-Lorenz Conference, University of Siena, Italy, May 2005)

More information

Information collection on a graph

Information collection on a graph Information collection on a grah Ilya O. Ryzhov Warren Powell October 25, 2009 Abstract We derive a knowledge gradient olicy for an otimal learning roblem on a grah, in which we use sequential measurements

More information

arxiv: v1 [physics.data-an] 26 Oct 2012

arxiv: v1 [physics.data-an] 26 Oct 2012 Constraints on Yield Parameters in Extended Maximum Likelihood Fits Till Moritz Karbach a, Maximilian Schlu b a TU Dortmund, Germany, moritz.karbach@cern.ch b TU Dortmund, Germany, maximilian.schlu@cern.ch

More information

John Weatherwax. Analysis of Parallel Depth First Search Algorithms

John Weatherwax. Analysis of Parallel Depth First Search Algorithms Sulementary Discussions and Solutions to Selected Problems in: Introduction to Parallel Comuting by Viin Kumar, Ananth Grama, Anshul Guta, & George Karyis John Weatherwax Chater 8 Analysis of Parallel

More information

Shadow Computing: An Energy-Aware Fault Tolerant Computing Model

Shadow Computing: An Energy-Aware Fault Tolerant Computing Model Shadow Comuting: An Energy-Aware Fault Tolerant Comuting Model Bryan Mills, Taieb Znati, Rami Melhem Deartment of Comuter Science University of Pittsburgh (bmills, znati, melhem)@cs.itt.edu Index Terms

More information

4. Score normalization technical details We now discuss the technical details of the score normalization method.

4. Score normalization technical details We now discuss the technical details of the score normalization method. SMT SCORING SYSTEM This document describes the scoring system for the Stanford Math Tournament We begin by giving an overview of the changes to scoring and a non-technical descrition of the scoring rules

More information

Developing A Deterioration Probabilistic Model for Rail Wear

Developing A Deterioration Probabilistic Model for Rail Wear International Journal of Traffic and Transortation Engineering 2012, 1(2): 13-18 DOI: 10.5923/j.ijtte.20120102.02 Develoing A Deterioration Probabilistic Model for Rail Wear Jabbar-Ali Zakeri *, Shahrbanoo

More information

Online Appendix to Accompany AComparisonof Traditional and Open-Access Appointment Scheduling Policies

Online Appendix to Accompany AComparisonof Traditional and Open-Access Appointment Scheduling Policies Online Aendix to Accomany AComarisonof Traditional and Oen-Access Aointment Scheduling Policies Lawrence W. Robinson Johnson Graduate School of Management Cornell University Ithaca, NY 14853-6201 lwr2@cornell.edu

More information

Estimation of component redundancy in optimal age maintenance

Estimation of component redundancy in optimal age maintenance EURO MAINTENANCE 2012, Belgrade 14-16 May 2012 Proceedings of the 21 st International Congress on Maintenance and Asset Management Estimation of comonent redundancy in otimal age maintenance Jorge ioa

More information

On the Relationship Between Packet Size and Router Performance for Heavy-Tailed Traffic 1

On the Relationship Between Packet Size and Router Performance for Heavy-Tailed Traffic 1 On the Relationshi Between Packet Size and Router Performance for Heavy-Tailed Traffic 1 Imad Antonios antoniosi1@southernct.edu CS Deartment MO117 Southern Connecticut State University 501 Crescent St.

More information

Algorithms for Air Traffic Flow Management under Stochastic Environments

Algorithms for Air Traffic Flow Management under Stochastic Environments Algorithms for Air Traffic Flow Management under Stochastic Environments Arnab Nilim and Laurent El Ghaoui Abstract A major ortion of the delay in the Air Traffic Management Systems (ATMS) in US arises

More information

Analysis of Multi-Hop Emergency Message Propagation in Vehicular Ad Hoc Networks

Analysis of Multi-Hop Emergency Message Propagation in Vehicular Ad Hoc Networks Analysis of Multi-Ho Emergency Message Proagation in Vehicular Ad Hoc Networks ABSTRACT Vehicular Ad Hoc Networks (VANETs) are attracting the attention of researchers, industry, and governments for their

More information

Analysis of some entrance probabilities for killed birth-death processes

Analysis of some entrance probabilities for killed birth-death processes Analysis of some entrance robabilities for killed birth-death rocesses Master s Thesis O.J.G. van der Velde Suervisor: Dr. F.M. Sieksma July 5, 207 Mathematical Institute, Leiden University Contents Introduction

More information

Power Aware Wireless File Downloading: A Constrained Restless Bandit Approach

Power Aware Wireless File Downloading: A Constrained Restless Bandit Approach PROC. WIOP 204 Power Aware Wireless File Downloading: A Constrained Restless Bandit Aroach Xiaohan Wei and Michael J. Neely, Senior Member, IEEE Abstract his aer treats ower-aware throughut maximization

More information

Hidden Predictors: A Factor Analysis Primer

Hidden Predictors: A Factor Analysis Primer Hidden Predictors: A Factor Analysis Primer Ryan C Sanchez Western Washington University Factor Analysis is a owerful statistical method in the modern research sychologist s toolbag When used roerly, factor

More information

q-ary Symmetric Channel for Large q

q-ary Symmetric Channel for Large q List-Message Passing Achieves Caacity on the q-ary Symmetric Channel for Large q Fan Zhang and Henry D Pfister Deartment of Electrical and Comuter Engineering, Texas A&M University {fanzhang,hfister}@tamuedu

More information

Characterizing the Behavior of a Probabilistic CMOS Switch Through Analytical Models and Its Verification Through Simulations

Characterizing the Behavior of a Probabilistic CMOS Switch Through Analytical Models and Its Verification Through Simulations Characterizing the Behavior of a Probabilistic CMOS Switch Through Analytical Models and Its Verification Through Simulations PINAR KORKMAZ, BILGE E. S. AKGUL and KRISHNA V. PALEM Georgia Institute of

More information

An Analysis of Reliable Classifiers through ROC Isometrics

An Analysis of Reliable Classifiers through ROC Isometrics An Analysis of Reliable Classifiers through ROC Isometrics Stijn Vanderlooy s.vanderlooy@cs.unimaas.nl Ida G. Srinkhuizen-Kuyer kuyer@cs.unimaas.nl Evgueni N. Smirnov smirnov@cs.unimaas.nl MICC-IKAT, Universiteit

More information

Preconditioning techniques for Newton s method for the incompressible Navier Stokes equations

Preconditioning techniques for Newton s method for the incompressible Navier Stokes equations Preconditioning techniques for Newton s method for the incomressible Navier Stokes equations H. C. ELMAN 1, D. LOGHIN 2 and A. J. WATHEN 3 1 Deartment of Comuter Science, University of Maryland, College

More information

Vision Graph Construction in Wireless Multimedia Sensor Networks

Vision Graph Construction in Wireless Multimedia Sensor Networks University of Nebraska - Lincoln DigitalCommons@University of Nebraska - Lincoln CSE Conference and Worksho Paers Comuter Science and Engineering, Deartment of 21 Vision Grah Construction in Wireless Multimedia

More information

An Improved Calibration Method for a Chopped Pyrgeometer

An Improved Calibration Method for a Chopped Pyrgeometer 96 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 17 An Imroved Calibration Method for a Choed Pyrgeometer FRIEDRICH FERGG OtoLab, Ingenieurbüro, Munich, Germany PETER WENDLING Deutsches Forschungszentrum

More information

A Qualitative Event-based Approach to Multiple Fault Diagnosis in Continuous Systems using Structural Model Decomposition

A Qualitative Event-based Approach to Multiple Fault Diagnosis in Continuous Systems using Structural Model Decomposition A Qualitative Event-based Aroach to Multile Fault Diagnosis in Continuous Systems using Structural Model Decomosition Matthew J. Daigle a,,, Anibal Bregon b,, Xenofon Koutsoukos c, Gautam Biswas c, Belarmino

More information

Efficient Approximations for Call Admission Control Performance Evaluations in Multi-Service Networks

Efficient Approximations for Call Admission Control Performance Evaluations in Multi-Service Networks Efficient Aroximations for Call Admission Control Performance Evaluations in Multi-Service Networks Emre A. Yavuz, and Victor C. M. Leung Deartment of Electrical and Comuter Engineering The University

More information

Amin, Osama; Abediseid, Walid; Alouini, Mohamed-Slim. Institute of Electrical and Electronics Engineers (IEEE)

Amin, Osama; Abediseid, Walid; Alouini, Mohamed-Slim. Institute of Electrical and Electronics Engineers (IEEE) KAUST Reository Outage erformance of cognitive radio systems with Imroer Gaussian signaling Item tye Authors Erint version DOI Publisher Journal Rights Conference Paer Amin Osama; Abediseid Walid; Alouini

More information

Delay characterization of multi-hop transmission in a Poisson field of interference

Delay characterization of multi-hop transmission in a Poisson field of interference 1 Delay characterization of multi-ho transmission in a Poisson field of interference Kostas Stamatiou and Martin Haenggi, Senior Member, IEEE Abstract We evaluate the end-to-end delay of a multi-ho transmission

More information

Outline. EECS150 - Digital Design Lecture 26 Error Correction Codes, Linear Feedback Shift Registers (LFSRs) Simple Error Detection Coding

Outline. EECS150 - Digital Design Lecture 26 Error Correction Codes, Linear Feedback Shift Registers (LFSRs) Simple Error Detection Coding Outline EECS150 - Digital Design Lecture 26 Error Correction Codes, Linear Feedback Shift Registers (LFSRs) Error detection using arity Hamming code for error detection/correction Linear Feedback Shift

More information

Cryptanalysis of Pseudorandom Generators

Cryptanalysis of Pseudorandom Generators CSE 206A: Lattice Algorithms and Alications Fall 2017 Crytanalysis of Pseudorandom Generators Instructor: Daniele Micciancio UCSD CSE As a motivating alication for the study of lattice in crytograhy we

More information

Approximating min-max k-clustering

Approximating min-max k-clustering Aroximating min-max k-clustering Asaf Levin July 24, 2007 Abstract We consider the roblems of set artitioning into k clusters with minimum total cost and minimum of the maximum cost of a cluster. The cost

More information

AI*IA 2003 Fusion of Multiple Pattern Classifiers PART III

AI*IA 2003 Fusion of Multiple Pattern Classifiers PART III AI*IA 23 Fusion of Multile Pattern Classifiers PART III AI*IA 23 Tutorial on Fusion of Multile Pattern Classifiers by F. Roli 49 Methods for fusing multile classifiers Methods for fusing multile classifiers

More information

Scaling Multiple Point Statistics for Non-Stationary Geostatistical Modeling

Scaling Multiple Point Statistics for Non-Stationary Geostatistical Modeling Scaling Multile Point Statistics or Non-Stationary Geostatistical Modeling Julián M. Ortiz, Steven Lyster and Clayton V. Deutsch Centre or Comutational Geostatistics Deartment o Civil & Environmental Engineering

More information

Statics and dynamics: some elementary concepts

Statics and dynamics: some elementary concepts 1 Statics and dynamics: some elementary concets Dynamics is the study of the movement through time of variables such as heartbeat, temerature, secies oulation, voltage, roduction, emloyment, rices and

More information

Numerical Linear Algebra

Numerical Linear Algebra Numerical Linear Algebra Numerous alications in statistics, articularly in the fitting of linear models. Notation and conventions: Elements of a matrix A are denoted by a ij, where i indexes the rows and

More information

Periodic scheduling 05/06/

Periodic scheduling 05/06/ Periodic scheduling T T or eriodic scheduling, the best that we can do is to design an algorithm which will always find a schedule if one exists. A scheduler is defined to be otimal iff it will find a

More information

Diverse Routing in Networks with Probabilistic Failures

Diverse Routing in Networks with Probabilistic Failures Diverse Routing in Networks with Probabilistic Failures Hyang-Won Lee, Member, IEEE, Eytan Modiano, Senior Member, IEEE, Kayi Lee, Member, IEEE Abstract We develo diverse routing schemes for dealing with

More information

The non-stochastic multi-armed bandit problem

The non-stochastic multi-armed bandit problem Submitted for journal ublication. The non-stochastic multi-armed bandit roblem Peter Auer Institute for Theoretical Comuter Science Graz University of Technology A-8010 Graz (Austria) auer@igi.tu-graz.ac.at

More information

Optimism, Delay and (In)Efficiency in a Stochastic Model of Bargaining

Optimism, Delay and (In)Efficiency in a Stochastic Model of Bargaining Otimism, Delay and In)Efficiency in a Stochastic Model of Bargaining Juan Ortner Boston University Setember 10, 2012 Abstract I study a bilateral bargaining game in which the size of the surlus follows

More information

Uncorrelated Multilinear Principal Component Analysis for Unsupervised Multilinear Subspace Learning

Uncorrelated Multilinear Principal Component Analysis for Unsupervised Multilinear Subspace Learning TNN-2009-P-1186.R2 1 Uncorrelated Multilinear Princial Comonent Analysis for Unsuervised Multilinear Subsace Learning Haiing Lu, K. N. Plataniotis and A. N. Venetsanooulos The Edward S. Rogers Sr. Deartment

More information

Analysis of execution time for parallel algorithm to dertmine if it is worth the effort to code and debug in parallel

Analysis of execution time for parallel algorithm to dertmine if it is worth the effort to code and debug in parallel Performance Analysis Introduction Analysis of execution time for arallel algorithm to dertmine if it is worth the effort to code and debug in arallel Understanding barriers to high erformance and redict

More information

A Simple Throughput Model for TCP Veno

A Simple Throughput Model for TCP Veno A Simle Throughut Model for TCP Veno Bin Zhou, Cheng Peng Fu, Dah-Ming Chiu, Chiew Tong Lau, and Lek Heng Ngoh School of Comuter Engineering, Nanyang Technological University, Singaore 639798 Email: {zhou00,

More information

New Schedulability Test Conditions for Non-preemptive Scheduling on Multiprocessor Platforms

New Schedulability Test Conditions for Non-preemptive Scheduling on Multiprocessor Platforms New Schedulability Test Conditions for Non-reemtive Scheduling on Multirocessor Platforms Technical Reort May 2008 Nan Guan 1, Wang Yi 2, Zonghua Gu 3 and Ge Yu 1 1 Northeastern University, Shenyang, China

More information

Distributed Rule-Based Inference in the Presence of Redundant Information

Distributed Rule-Based Inference in the Presence of Redundant Information istribution Statement : roved for ublic release; distribution is unlimited. istributed Rule-ased Inference in the Presence of Redundant Information June 8, 004 William J. Farrell III Lockheed Martin dvanced

More information

Hotelling s Two- Sample T 2

Hotelling s Two- Sample T 2 Chater 600 Hotelling s Two- Samle T Introduction This module calculates ower for the Hotelling s two-grou, T-squared (T) test statistic. Hotelling s T is an extension of the univariate two-samle t-test

More information

ECE 534 Information Theory - Midterm 2

ECE 534 Information Theory - Midterm 2 ECE 534 Information Theory - Midterm Nov.4, 009. 3:30-4:45 in LH03. You will be given the full class time: 75 minutes. Use it wisely! Many of the roblems have short answers; try to find shortcuts. You

More information

Modeling and Estimation of Full-Chip Leakage Current Considering Within-Die Correlation

Modeling and Estimation of Full-Chip Leakage Current Considering Within-Die Correlation 6.3 Modeling and Estimation of Full-Chi Leaage Current Considering Within-Die Correlation Khaled R. eloue, Navid Azizi, Farid N. Najm Deartment of ECE, University of Toronto,Toronto, Ontario, Canada {haled,nazizi,najm}@eecg.utoronto.ca

More information

Availability and Maintainability. Piero Baraldi

Availability and Maintainability. Piero Baraldi Availability and Maintainability 1 Introduction: reliability and availability System tyes Non maintained systems: they cannot be reaired after a failure (a telecommunication satellite, a F1 engine, a vessel

More information

SIMULATED ANNEALING AND JOINT MANUFACTURING BATCH-SIZING. Ruhul SARKER. Xin YAO

SIMULATED ANNEALING AND JOINT MANUFACTURING BATCH-SIZING. Ruhul SARKER. Xin YAO Yugoslav Journal of Oerations Research 13 (003), Number, 45-59 SIMULATED ANNEALING AND JOINT MANUFACTURING BATCH-SIZING Ruhul SARKER School of Comuter Science, The University of New South Wales, ADFA,

More information

Analyses of Orthogonal and Non-Orthogonal Steering Vectors at Millimeter Wave Systems

Analyses of Orthogonal and Non-Orthogonal Steering Vectors at Millimeter Wave Systems Analyses of Orthogonal and Non-Orthogonal Steering Vectors at Millimeter Wave Systems Hsiao-Lan Chiang, Tobias Kadur, and Gerhard Fettweis Vodafone Chair for Mobile Communications Technische Universität

More information

Proof: We follow thearoach develoed in [4]. We adot a useful but non-intuitive notion of time; a bin with z balls at time t receives its next ball at

Proof: We follow thearoach develoed in [4]. We adot a useful but non-intuitive notion of time; a bin with z balls at time t receives its next ball at A Scaling Result for Exlosive Processes M. Mitzenmacher Λ J. Sencer We consider the following balls and bins model, as described in [, 4]. Balls are sequentially thrown into bins so that the robability

More information

CHAPTER-II Control Charts for Fraction Nonconforming using m-of-m Runs Rules

CHAPTER-II Control Charts for Fraction Nonconforming using m-of-m Runs Rules CHAPTER-II Control Charts for Fraction Nonconforming using m-of-m Runs Rules. Introduction: The is widely used in industry to monitor the number of fraction nonconforming units. A nonconforming unit is

More information

Sets of Real Numbers

Sets of Real Numbers Chater 4 Sets of Real Numbers 4. The Integers Z and their Proerties In our revious discussions about sets and functions the set of integers Z served as a key examle. Its ubiquitousness comes from the fact

More information

Adaptive estimation with change detection for streaming data

Adaptive estimation with change detection for streaming data Adative estimation with change detection for streaming data A thesis resented for the degree of Doctor of Philosohy of the University of London and the Diloma of Imerial College by Dean Adam Bodenham Deartment

More information

Minimax Design of Nonnegative Finite Impulse Response Filters

Minimax Design of Nonnegative Finite Impulse Response Filters Minimax Design of Nonnegative Finite Imulse Resonse Filters Xiaoing Lai, Anke Xue Institute of Information and Control Hangzhou Dianzi University Hangzhou, 3118 China e-mail: laix@hdu.edu.cn; akxue@hdu.edu.cn

More information

State Estimation with ARMarkov Models

State Estimation with ARMarkov Models Deartment of Mechanical and Aerosace Engineering Technical Reort No. 3046, October 1998. Princeton University, Princeton, NJ. State Estimation with ARMarkov Models Ryoung K. Lim 1 Columbia University,

More information

An Ant Colony Optimization Approach to the Probabilistic Traveling Salesman Problem

An Ant Colony Optimization Approach to the Probabilistic Traveling Salesman Problem An Ant Colony Otimization Aroach to the Probabilistic Traveling Salesman Problem Leonora Bianchi 1, Luca Maria Gambardella 1, and Marco Dorigo 2 1 IDSIA, Strada Cantonale Galleria 2, CH-6928 Manno, Switzerland

More information

Asymptotically Optimal Simulation Allocation under Dependent Sampling

Asymptotically Optimal Simulation Allocation under Dependent Sampling Asymtotically Otimal Simulation Allocation under Deendent Samling Xiaoing Xiong The Robert H. Smith School of Business, University of Maryland, College Park, MD 20742-1815, USA, xiaoingx@yahoo.com Sandee

More information

#A64 INTEGERS 18 (2018) APPLYING MODULAR ARITHMETIC TO DIOPHANTINE EQUATIONS

#A64 INTEGERS 18 (2018) APPLYING MODULAR ARITHMETIC TO DIOPHANTINE EQUATIONS #A64 INTEGERS 18 (2018) APPLYING MODULAR ARITHMETIC TO DIOPHANTINE EQUATIONS Ramy F. Taki ElDin Physics and Engineering Mathematics Deartment, Faculty of Engineering, Ain Shams University, Cairo, Egyt

More information

On the capacity of the general trapdoor channel with feedback

On the capacity of the general trapdoor channel with feedback On the caacity of the general tradoor channel with feedback Jui Wu and Achilleas Anastasooulos Electrical Engineering and Comuter Science Deartment University of Michigan Ann Arbor, MI, 48109-1 email:

More information

Elementary Analysis in Q p

Elementary Analysis in Q p Elementary Analysis in Q Hannah Hutter, May Szedlák, Phili Wirth November 17, 2011 This reort follows very closely the book of Svetlana Katok 1. 1 Sequences and Series In this section we will see some

More information

Factors Effect on the Saturation Parameter S and there Influences on the Gain Behavior of Ytterbium Doped Fiber Amplifier

Factors Effect on the Saturation Parameter S and there Influences on the Gain Behavior of Ytterbium Doped Fiber Amplifier Australian Journal of Basic and Alied Sciences, 5(12): 2010-2020, 2011 ISSN 1991-8178 Factors Effect on the Saturation Parameter S and there Influences on the Gain Behavior of Ytterbium Doed Fiber Amlifier

More information

Plotting the Wilson distribution

Plotting the Wilson distribution , Survey of English Usage, University College London Setember 018 1 1. Introduction We have discussed the Wilson score interval at length elsewhere (Wallis 013a, b). Given an observed Binomial roortion

More information

1 Gambler s Ruin Problem

1 Gambler s Ruin Problem Coyright c 2017 by Karl Sigman 1 Gambler s Ruin Problem Let N 2 be an integer and let 1 i N 1. Consider a gambler who starts with an initial fortune of $i and then on each successive gamble either wins

More information

Anytime communication over the Gilbert-Eliot channel with noiseless feedback

Anytime communication over the Gilbert-Eliot channel with noiseless feedback Anytime communication over the Gilbert-Eliot channel with noiseless feedback Anant Sahai, Salman Avestimehr, Paolo Minero Deartment of Electrical Engineering and Comuter Sciences University of California

More information

On Code Design for Simultaneous Energy and Information Transfer

On Code Design for Simultaneous Energy and Information Transfer On Code Design for Simultaneous Energy and Information Transfer Anshoo Tandon Electrical and Comuter Engineering National University of Singaore Email: anshoo@nus.edu.sg Mehul Motani Electrical and Comuter

More information

1. INTRODUCTION. Fn 2 = F j F j+1 (1.1)

1. INTRODUCTION. Fn 2 = F j F j+1 (1.1) CERTAIN CLASSES OF FINITE SUMS THAT INVOLVE GENERALIZED FIBONACCI AND LUCAS NUMBERS The beautiful identity R.S. Melham Deartment of Mathematical Sciences, University of Technology, Sydney PO Box 23, Broadway,

More information

Statistical Multiplexing Gain of Link Scheduling Algorithms in QoS Networks (Short Version)

Statistical Multiplexing Gain of Link Scheduling Algorithms in QoS Networks (Short Version) 1 Statistical Multilexing Gain of Link Scheduling Algorithms in QoS Networks (Short Version) Technical Reort: University of Virginia, CS-99-23, July 1999 Robert Boorstyn Almut Burchard Jörg Liebeherr y

More information

Computer arithmetic. Intensive Computation. Annalisa Massini 2017/2018

Computer arithmetic. Intensive Computation. Annalisa Massini 2017/2018 Comuter arithmetic Intensive Comutation Annalisa Massini 7/8 Intensive Comutation - 7/8 References Comuter Architecture - A Quantitative Aroach Hennessy Patterson Aendix J Intensive Comutation - 7/8 3

More information

RUN-TO-RUN CONTROL AND PERFORMANCE MONITORING OF OVERLAY IN SEMICONDUCTOR MANUFACTURING. 3 Department of Chemical Engineering

RUN-TO-RUN CONTROL AND PERFORMANCE MONITORING OF OVERLAY IN SEMICONDUCTOR MANUFACTURING. 3 Department of Chemical Engineering Coyright 2002 IFAC 15th Triennial World Congress, Barcelona, Sain RUN-TO-RUN CONTROL AND PERFORMANCE MONITORING OF OVERLAY IN SEMICONDUCTOR MANUFACTURING C.A. Bode 1, B.S. Ko 2, and T.F. Edgar 3 1 Advanced

More information

On the Chvatál-Complexity of Knapsack Problems

On the Chvatál-Complexity of Knapsack Problems R u t c o r Research R e o r t On the Chvatál-Comlexity of Knasack Problems Gergely Kovács a Béla Vizvári b RRR 5-08, October 008 RUTCOR Rutgers Center for Oerations Research Rutgers University 640 Bartholomew

More information

The Noise Power Ratio - Theory and ADC Testing

The Noise Power Ratio - Theory and ADC Testing The Noise Power Ratio - Theory and ADC Testing FH Irons, KJ Riley, and DM Hummels Abstract This aer develos theory behind the noise ower ratio (NPR) testing of ADCs. A mid-riser formulation is used for

More information

Multi-Operation Multi-Machine Scheduling

Multi-Operation Multi-Machine Scheduling Multi-Oeration Multi-Machine Scheduling Weizhen Mao he College of William and Mary, Williamsburg VA 3185, USA Abstract. In the multi-oeration scheduling that arises in industrial engineering, each job

More information

WIRELESS energy transfer (WET), where receivers harvest

WIRELESS energy transfer (WET), where receivers harvest 6898 IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 15, NO. 10, OCTOBER 2016 User-Centric Energy Efficiency Maximization for Wireless Powered Communications Qingqing Wu, Student Member, IEEE, Wen Chen,

More information

An Investigation on the Numerical Ill-conditioning of Hybrid State Estimators

An Investigation on the Numerical Ill-conditioning of Hybrid State Estimators An Investigation on the Numerical Ill-conditioning of Hybrid State Estimators S. K. Mallik, Student Member, IEEE, S. Chakrabarti, Senior Member, IEEE, S. N. Singh, Senior Member, IEEE Deartment of Electrical

More information

Lower bound solutions for bearing capacity of jointed rock

Lower bound solutions for bearing capacity of jointed rock Comuters and Geotechnics 31 (2004) 23 36 www.elsevier.com/locate/comgeo Lower bound solutions for bearing caacity of jointed rock D.J. Sutcliffe a, H.S. Yu b, *, S.W. Sloan c a Deartment of Civil, Surveying

More information

Lower Confidence Bound for Process-Yield Index S pk with Autocorrelated Process Data

Lower Confidence Bound for Process-Yield Index S pk with Autocorrelated Process Data Quality Technology & Quantitative Management Vol. 1, No.,. 51-65, 15 QTQM IAQM 15 Lower onfidence Bound for Process-Yield Index with Autocorrelated Process Data Fu-Kwun Wang * and Yeneneh Tamirat Deartment

More information

Radial Basis Function Networks: Algorithms

Radial Basis Function Networks: Algorithms Radial Basis Function Networks: Algorithms Introduction to Neural Networks : Lecture 13 John A. Bullinaria, 2004 1. The RBF Maing 2. The RBF Network Architecture 3. Comutational Power of RBF Networks 4.

More information

arxiv:cond-mat/ v2 25 Sep 2002

arxiv:cond-mat/ v2 25 Sep 2002 Energy fluctuations at the multicritical oint in two-dimensional sin glasses arxiv:cond-mat/0207694 v2 25 Se 2002 1. Introduction Hidetoshi Nishimori, Cyril Falvo and Yukiyasu Ozeki Deartment of Physics,

More information

Tests for Two Proportions in a Stratified Design (Cochran/Mantel-Haenszel Test)

Tests for Two Proportions in a Stratified Design (Cochran/Mantel-Haenszel Test) Chater 225 Tests for Two Proortions in a Stratified Design (Cochran/Mantel-Haenszel Test) Introduction In a stratified design, the subects are selected from two or more strata which are formed from imortant

More information

KEY ISSUES IN THE ANALYSIS OF PILES IN LIQUEFYING SOILS

KEY ISSUES IN THE ANALYSIS OF PILES IN LIQUEFYING SOILS 4 th International Conference on Earthquake Geotechnical Engineering June 2-28, 27 KEY ISSUES IN THE ANALYSIS OF PILES IN LIQUEFYING SOILS Misko CUBRINOVSKI 1, Hayden BOWEN 1 ABSTRACT Two methods for analysis

More information

A STUDY ON THE UTILIZATION OF COMPATIBILITY METRIC IN THE AHP: APPLYING TO SOFTWARE PROCESS ASSESSMENTS

A STUDY ON THE UTILIZATION OF COMPATIBILITY METRIC IN THE AHP: APPLYING TO SOFTWARE PROCESS ASSESSMENTS ISAHP 2005, Honolulu, Hawaii, July 8-10, 2003 A SUDY ON HE UILIZAION OF COMPAIBILIY MERIC IN HE AHP: APPLYING O SOFWARE PROCESS ASSESSMENS Min-Suk Yoon Yosu National University San 96-1 Dundeok-dong Yeosu

More information

Supplementary Materials for Robust Estimation of the False Discovery Rate

Supplementary Materials for Robust Estimation of the False Discovery Rate Sulementary Materials for Robust Estimation of the False Discovery Rate Stan Pounds and Cheng Cheng This sulemental contains roofs regarding theoretical roerties of the roosed method (Section S1), rovides

More information

End-to-End Delay Minimization in Thermally Constrained Distributed Systems

End-to-End Delay Minimization in Thermally Constrained Distributed Systems End-to-End Delay Minimization in Thermally Constrained Distributed Systems Pratyush Kumar, Lothar Thiele Comuter Engineering and Networks Laboratory (TIK) ETH Zürich, Switzerland {ratyush.kumar, lothar.thiele}@tik.ee.ethz.ch

More information

Mobility-Induced Service Migration in Mobile. Micro-Clouds

Mobility-Induced Service Migration in Mobile. Micro-Clouds arxiv:503054v [csdc] 7 Mar 205 Mobility-Induced Service Migration in Mobile Micro-Clouds Shiiang Wang, Rahul Urgaonkar, Ting He, Murtaza Zafer, Kevin Chan, and Kin K LeungTime Oerating after ossible Deartment

More information

Morten Frydenberg Section for Biostatistics Version :Friday, 05 September 2014

Morten Frydenberg Section for Biostatistics Version :Friday, 05 September 2014 Morten Frydenberg Section for Biostatistics Version :Friday, 05 Setember 204 All models are aroximations! The best model does not exist! Comlicated models needs a lot of data. lower your ambitions or get

More information